viable/strict/1791906918: [AARCH64] Use oneDNN jit:int8 matmul for torch._int_mm (#200389)
- PyTorch: 2261 events in the last 90 days
- PyTorch: 2231th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1791905525
What happened
On AArch64, torch._int_mm always uses the native fallback loop. TorchAO's Int8DynamicActivationInt8WeightConfig calls _int_mm on CPU, so int8 linear layers don't benefit from oneDNN. Since v3.13, oneDNN's AArch64 jit:int8 matmul (SVE + I8MM) supports s32 output ( uxlfoundation/oneDNN#5267 ). This PR routes _int_mm to it. Changes (AArch64 only; x86 unchanged) Dispatch _int_mm to oneDNN when the CPU has SVE and I8MM.…
Summary assembled by rule from the sources below